complex threading with OpenMP

Question

I need to switch from boost::thread to OpenMP because boss says so.

The problem is quiet simple: the result of a simulation is written to disk every 5 iteration (int it = 5,10,15...). For the sake of simplicity, suppose I have an 8-core CPU. I created 9 threads; thread 0 is used for IO, other 8 for computation. When (it%5 == 0), I check thread 0 to see if it has finished. If yes, I create another thread, call 0, and ask it to write the result to disk. If not, all threads have to wait. Usually, the time it takes to write out a result is less than 5 iterations, so I effectively "hide" the IO cost.

I have spent a few hours looking into OpenMP and I guess the same algorithm can be done with the "task" construct but I don't see how I can synchronize the threads. OpenMP experts please help. Thanks.

The current pseudo code look like this

boost::thread pool[9];
for(int it=0;it<1000;it++)
{
    - simulate using pool[1,8]
    - if(it%5 == 0)
         + check pool[0]
         + if finished: create new thread, assign to pool[0], write data out
         + if not, wait
}

I love the "boss says so" educate the Boss, OpenMP is NOT threads, for that you may look at Intel TBB. — whatnick, Jun 29 '11 at 23:35
@whatnick: though it has a thread class, Intel TBB is not about threads either. — Alexey Kukanov, Jun 30 '11 at 20:11
OpenMP has "lock" (similar to mutex) that I think one can use to track threads. In this case, I think the solution is to have the IO task unlock the lock, and to make sure the lock is unlocked before spawning another IO task. When I asked this question, I did not know of "locks". I think I have an answer to my own question; but I still need OpenMP experts to verify. — JohnC, Jul 03 '11 at 03:36

score 2 · Answer 1 · answered Jun 29 '11 at 23:37

2

Intel has a very nice answer to the OpenMP vs Threads dilemma, I would defer to them and ask your boss to get some education.

OpenMP is very loop oriented, you parallize the entire loop rather than have synchronised threads.

answered Jun 29 '11 at 23:37

whatnick

5,400
3
19
35

educating boss is out of the question. Same with boost, which is nice but boss doesn't want to package the library with our software. OpenMP 3 has a new "task" construct to spawn threads at will so there must be some way to know when the thread has finished... – JohnC Jun 30 '11 at 03:56
Intel libraries... not too fond of. Wait until you have to read the manual in Rusnglish (i am talking about MKL library) then you see what i mean. My company sticks with gcc, and I think boss is right about this... – JohnC Jun 30 '11 at 04:00

score 0 · Answer 2 · answered Jun 30 '11 at 20:29

Your design seems right to me overall: having a separate thread for I/O and a thread pool for computations is right. You might possibly replace Boost threads in pool[1..8] with OpenMP for the computational part, but I would not go beyond that. If you can't use Boost, use POSIX threads.

complex threading with OpenMP

2 Answers2